Variations in the Incidence of Postpartum Hemorrhage Across Hospitals in California

Objective: Because postpartum hemorrhage may result from factors related to obstetrical practice patterns, we examined the variability of postpartum hemorrhage and related risk factors (obstetrical trauma, chorioamnionitis, and protracted labor) across hospital types and hospitals in California. Methods: Linked birth certificate and hospital discharge data from 507,410 births in California in 1997 were analyzed. Cases were identified using International Classification of Diseases, Ninth Edition, Clinical Modification (ICD-9-CM) codes. Comparisons were made across hospital types and individual hospitals. Risk adjustments were made using 1) sample restriction to a subset of 324,671 low-risk women, and 2) Bayesian hierarchical logistic regression model to simultaneously quantify the effects of patient-level and hospital-level risk factors. Results: Postpartum hemorrhage complicated 2.4% of live births. The incidence ranged from 1.6% for corporate hospitals to 4.9% for university hospitals in the full sample, and from 1.4% for corporate hospitals to 3.9% for university hospitals in the low-risk sample. Low-risk women who delivered at government, HMO and university hospitals had two- to threefold increased odds (odds ratios 1.98 to 2.71; 95% confidence sets ranged from 1.52 to 4.62) of having postpartum hemorrhage compared to women who delivered at corporate hospitals, irrespective of patient-level characteristics. They also had significantly higher rates of obstetrical trauma and chorioamnionitis. Greater variations were observed across individual hospitals. Conclusion: The incidence of postpartum hemorrhage and related risk factors varied substantially across hospital types and hospitals in California. Further studies using primary data sources are needed to determine whether these variations are related to the processesof care.

Appendices

APPENDIX A: MATERNAL, FETAL, OR PLACENTAL CONDITIONS EXCLUDED FROM THE LOW-RISK SAMPLE

• Asthma

• Congenital fetal CNS anomaly or chromosomal abnormality

• Diabetes or abnormal glucose tolerance

• Herpes

• Intrauterine fetal demise

• Intrauterine growth restriction

• Macrosomia

• Malpresentation

• Maternal cerebral hemorrhage

• Maternal congenital and other heart disease

• Maternal hypotension or obstetrical shock

• Maternal renal abnormalities

• Maternal soft tissue disorder

• Maternal thyroid abnormalities

• Mental disorder

• Multiple gestation

• Oligohydramnios

• Other types of hypertension

• Polyhyramnios

• Preterm gestation

• Pulmonary embolism

• Severe hypertension: severe pre-eclampsia or eclampsia

• Uterine rupture

APPENDIX B: BAYESIAN HIERARCHICAL LOGISTIC REGRESSION MODEL USED TO ESTIMATE THE ADJUSTED RATES OF POSTPARTUM HEMORRHAGE, OBSTETRICAL TRAUMA, CHORIOAMNIONITIS, AND PROTRACTED LABOR FOR INDIVIDUAL HOSPITALS

The probability of each of the four events (postpartum hemorrhage, protracted labor, chorioamnionitis, and obstetrical trauma) was modeled using hierarchical logistic regression models. This modeling framework facilitates the analysis of provider variations in outcomes by separating provider-related factors from patient-related ones. The model consists of two equations that are simultaneously estimated, one (patient level) to model within hospital variability, and the other (hospital level) to model between hospital variability. For each dichotomous outcome, the hierarchical logistic regression equations are as follows:

P(Y _ij = 1) = P _ij is the Bernoulli probability of an event for patient i in hospital j.

• 1. Patient level equation:

  Logit (P _ij ) = β _{1 j} + β _{2 j} (Age ≥ 35) _ij + β _{3 j} (Poor PNC) _ij + β _{4 j} (Asian) _ij + β _{5 j} (African American) _ij + β _{7 j} (Hispanic) _ij + β _{8 j} (Native American) _ij + β _{9 j} (Nulliparous) _ij + β _{10 j} (Prior CS) _ij

  Covariates include age, race-ethnicity, parity, prenatal care utilization (10), and prior cesarean delivery. The reference group consists of Caucasian women less than 35 years of age, with adequate or adequate-plus prenatal care, multiparous with no prior cesarean delivery. The beta coefficients from this logistic regression (log-odds scale) measure the effect of patient characteristics on outcome within hospital j. The intercept β _{1 j} is the baseline (for the reference group) event rate for hospital j and the other β slopes measure the effects of the covariates on outcome.

• 2. Hospital level equation:

  β _j is distributed multivariate normal with mean μ and variance–covariance matrix D (bold parameters are multi-dimensional) where,

  $$\mu = \gamma _1 + \gamma _2 ({\rm Government})_j + \gamma _3 ({\rm District})_j\\\quad +\, \gamma _4 ({\rm HMO})_j + \gamma _5 ({\rm Not}\;{\rm for}\;{\rm profit})_j \\ \quad +\,\gamma _6 ({\rm UC})_j + \gamma _7 ({\rm Other}\;{\rm hospital}\;{\rm covariates}\;\\\qquad {\rm such}\; {\rm as}\;{\rm OB}\;{\rm teaching}\;{\rm status})_j$$

Covariates for the hospital-level equation include indicator variables for hospital ownership and OB teaching status. The intercept γ ₁ is a reference parameter representing corporate hospitals (none of which are obstetrics teaching hospitals).

We assume that the hospital-specific parameter β _j is a multivariate normal variate with mean μ that is linearly related to the hospital characteristics. The components of μ are therefore hospital class expectations given the class characteristics. The γ's are parameters common to all hospitals, measuring the effect of hospital characteristics on the expected hospital class rate of outcome. In other words, the components of γ link between hospital characteristics and hospital-specific effects. While the β _j are directly related to the patient covariates, the γ's are directly related to hospital level covariates. Note that not only mean hospital rate of outcomes (β intercepts) may vary by hospital characteristics but also β slopes are allowed to vary by hospital characteristics.

The model was estimated under a Bayesian framework that assigns prior probability distributions to all unknown parameters (specifically to γ and D). Posterior distributions of the model parameters (conditional on the data) were derived using Markov Chain Monte Carlo methodology. For each outcome a single Gibbs sampler string was used for inference, implemented using the WinBUGS software (5), with a burn-in of 1000 iterations and a further 2000 iteration used for inference. Starting values for parameters were calculated using logistic regression models. Estimated posterior means for odd ratios are reported with significance determined using 95% confidence sets.